Datasets:

WINGNUS
/

ACL-OCL

	{
	"paper_id": "I11-1002",
	"header": {
	"generated_with": "S2ORC 1.0.0",
	"date_generated": "2023-01-19T07:31:04.064691Z"
	},
	"title": "Dependency-directed Tree Kernel-based Protein-Protein Interaction Extraction from Biomedical Literature",
	"authors": [
	{
	"first": "Longhua",
	"middle": [],
	"last": "Qian",
	"suffix": "",
	"affiliation": {
	"laboratory": "Natural Language Processing Lab",
	"institution": "Soochow University",
	"location": {
	"addrLine": "1 Shizi Street",
	"postCode": "215006",
	"settlement": "Suzhou",
	"country": "China"
	}
	},
	"email": "qianlonghua@suda.edu.cn"
	},
	{
	"first": "Guodong",
	"middle": [],
	"last": "Zhou",
	"suffix": "",
	"affiliation": {
	"laboratory": "Natural Language Processing Lab",
	"institution": "Soochow University",
	"location": {
	"addrLine": "1 Shizi Street",
	"postCode": "215006",
	"settlement": "Suzhou",
	"country": "China"
	}
	},
	"email": "gdzhou@suda.edu.cn"
	}
	],
	"year": "",
	"venue": null,
	"identifiers": {},
	"abstract": "Structured information plays a critical role in many NLP tasks, such as semantic relation extraction between named entities and semantic role labeling. This paper proposes a principled way to automatically generate constituent structure representation for tree kernel-based protein-protein interaction (PPI) extraction. The main idea behind our approach is that the critical portion in a constituent parse tree for PPI extraction can be automatically determined by the shortest dependency path between the two involved proteins, while other portion can be regarded as noise and ignored safely. Evaluation on multiple PPI corpora shows that our dependency-directed tree kernel-based method achieves promising results. This justifies the effectiveness of tree kernelbased methods for PPI extraction, in particular the advantage of dependency-directed constituent structure representation.",
	"pdf_parse": {
	"paper_id": "I11-1002",
	"_pdf_hash": "",
	"abstract": [
	{
	"text": "Structured information plays a critical role in many NLP tasks, such as semantic relation extraction between named entities and semantic role labeling. This paper proposes a principled way to automatically generate constituent structure representation for tree kernel-based protein-protein interaction (PPI) extraction. The main idea behind our approach is that the critical portion in a constituent parse tree for PPI extraction can be automatically determined by the shortest dependency path between the two involved proteins, while other portion can be regarded as noise and ignored safely. Evaluation on multiple PPI corpora shows that our dependency-directed tree kernel-based method achieves promising results. This justifies the effectiveness of tree kernelbased methods for PPI extraction, in particular the advantage of dependency-directed constituent structure representation.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Abstract",
	"sec_num": null
	}
	],
	"body_text": [
	{
	"text": "Since determining protein interaction partners is crucial to understand both the functional role of individual proteins and the organization of the entire biological process, there is a significant interest in protein-protein interaction (PPI) extraction. However, manual collection of relevant PPI information from thousands of biomedical research papers published every day (e.g. MED-LINE) is so time-consuming and labordemanding that automatic extraction approaches with the help of NLP techniques become necessary.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "In principle, PPI extraction is much like the semantic relation extraction subtask (so called Relation Detection and Classification, RDC) defined by the ACE project (ACE, 2002 (ACE, -2007 in the newswire domain. Therefore, various kinds of machine learning methods have been borrowed from the newswire domain to the biomedical domain: feature-based methods (Mitsumori et al., 2006; Giuliano et al., 2006; Saetre et al., 2007; Liu et al., 2010) and kernelbased methods (Bunescu et al., 2005a; Erkan et al., 2007; Airola et al., 2008; Kim et al., 2010) .",
	"cite_spans": [
	{
	"start": 165,
	"end": 175,
	"text": "(ACE, 2002",
	"ref_id": null
	},
	{
	"start": 176,
	"end": 187,
	"text": "(ACE, -2007",
	"ref_id": null
	},
	{
	"start": 357,
	"end": 381,
	"text": "(Mitsumori et al., 2006;",
	"ref_id": "BIBREF18"
	},
	{
	"start": 382,
	"end": 404,
	"text": "Giuliano et al., 2006;",
	"ref_id": "BIBREF12"
	},
	{
	"start": 405,
	"end": 425,
	"text": "Saetre et al., 2007;",
	"ref_id": "BIBREF28"
	},
	{
	"start": 426,
	"end": 443,
	"text": "Liu et al., 2010)",
	"ref_id": "BIBREF17"
	},
	{
	"start": 468,
	"end": 491,
	"text": "(Bunescu et al., 2005a;",
	"ref_id": "BIBREF2"
	},
	{
	"start": 492,
	"end": 511,
	"text": "Erkan et al., 2007;",
	"ref_id": "BIBREF10"
	},
	{
	"start": 512,
	"end": 532,
	"text": "Airola et al., 2008;",
	"ref_id": "BIBREF1"
	},
	{
	"start": 533,
	"end": 550,
	"text": "Kim et al., 2010)",
	"ref_id": "BIBREF15"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Early studies on PPI extraction employ feature-based methods. However, the featurebased methods often fail to effectively capture the structured information, which is essential to identify the relationship between two proteins in a constituent or dependency-based syntactic representation.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "With the wide adoption of kernel-based methods to many NLP tasks, particularly for semantic relation extraction and semantic role labeling, various kernels such as subsequence kernels (Bunescu et al., 2005a) and tree kernels (Li et al., 2008) have been applied to PPI extraction. On one hand, dependency-based kernels, such as edit distance kernels (Erkan et al. 2007) , graph kernels (Airola et al., 2008) and subsequence kernels (Kim et al., 2010) , show some promising results for PPI extraction. This suggests that dependency information plays a critical role in PPI extraction, much like semantic relation extraction in the newswire narratives (Culotta and Sorensen, 2004; Bunescu et al., 2005b) . On the other hand, while tree kernels based on constituent parse trees achieve great success in semantic relation extraction (Zhang et al., 2006; Zhou et al., 2007a; Qian et al., 2008) and semantic role labeling (Moschitti, 2004; Zhang et al., 2008) from the newswire narratives, they haven't been fully explored for PPI extraction in the biomedical domain. Considering the similarity between the task of PPI extraction from the biomedical domain and that of relation extraction from the newswire domain, one question naturally arises: \"How can kernel-based PPI extraction benefit from the constituent parse tree structure?\"",
	"cite_spans": [
	{
	"start": 184,
	"end": 207,
	"text": "(Bunescu et al., 2005a)",
	"ref_id": "BIBREF2"
	},
	{
	"start": 225,
	"end": 242,
	"text": "(Li et al., 2008)",
	"ref_id": "BIBREF16"
	},
	{
	"start": 349,
	"end": 368,
	"text": "(Erkan et al. 2007)",
	"ref_id": "BIBREF10"
	},
	{
	"start": 385,
	"end": 406,
	"text": "(Airola et al., 2008)",
	"ref_id": "BIBREF1"
	},
	{
	"start": 431,
	"end": 449,
	"text": "(Kim et al., 2010)",
	"ref_id": "BIBREF15"
	},
	{
	"start": 649,
	"end": 677,
	"text": "(Culotta and Sorensen, 2004;",
	"ref_id": "BIBREF7"
	},
	{
	"start": 678,
	"end": 700,
	"text": "Bunescu et al., 2005b)",
	"ref_id": "BIBREF3"
	},
	{
	"start": 828,
	"end": 848,
	"text": "(Zhang et al., 2006;",
	"ref_id": "BIBREF31"
	},
	{
	"start": 849,
	"end": 868,
	"text": "Zhou et al., 2007a;",
	"ref_id": "BIBREF35"
	},
	{
	"start": 869,
	"end": 887,
	"text": "Qian et al., 2008)",
	"ref_id": "BIBREF26"
	},
	{
	"start": 915,
	"end": 932,
	"text": "(Moschitti, 2004;",
	"ref_id": "BIBREF21"
	},
	{
	"start": 933,
	"end": 952,
	"text": "Zhang et al., 2008)",
	"ref_id": "BIBREF32"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "To address this question, this paper presents a principled way to automatically generate a precise and concise constituent parse tree representation for kernel-based methods, motivated by the success of employing dependency information in PPI extraction. This is done by taking advantage of the shortest dependency path between two involved proteins in the dependency parse tree structure of a sentence. Specifically, only the words appearing on the shortest dependency path and their associated constituents in the constituent parse tree are considered as necessary and thus kept as the essential part of the constituent parse tree. In this paper, we refer to it as SDP-CPT (Shortest Dependency Path-directed Constituent Parse Tree). Experimental results on several major PPI corpora show the effectiveness of dependency-directed constituent structure representation and its preference over other state-ofthe-art structure representations.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "The rest of this paper is organized as follows. First, related work in PPI extraction is overviewed in Section 2. Then, Section 3 elaborates our shortest dependency path-directed constituent parse tree structure. Section 4 reports the experimental results on major PPI corpora. Finally we conclude our work in Section 5.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Due to space limitation, this section only gives an overview on kernel-based methods on PPI extraction in the biomedical domain as well as semantic relation extraction in the newswire domain. For details about feature-based methods, please refer to related studies in the biomedical domain (Mitsumori et al., 2006; Giuliano et al., 2006; Liu et al., 2010) and those in the newswire domain (Zhao et al., 2005; Zhou et al. 2005 Zhou et al. , 2007b , respectively.",
	"cite_spans": [
	{
	"start": 290,
	"end": 314,
	"text": "(Mitsumori et al., 2006;",
	"ref_id": "BIBREF18"
	},
	{
	"start": 315,
	"end": 337,
	"text": "Giuliano et al., 2006;",
	"ref_id": "BIBREF12"
	},
	{
	"start": 338,
	"end": 355,
	"text": "Liu et al., 2010)",
	"ref_id": "BIBREF17"
	},
	{
	"start": 389,
	"end": 408,
	"text": "(Zhao et al., 2005;",
	"ref_id": "BIBREF33"
	},
	{
	"start": 409,
	"end": 425,
	"text": "Zhou et al. 2005",
	"ref_id": "BIBREF34"
	},
	{
	"start": 426,
	"end": 445,
	"text": "Zhou et al. , 2007b",
	"ref_id": "BIBREF36"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "Representative kernel-based methods on PPI extraction take advantage of lexical or dependency information. Bunescu et al. (2005a) adopt a generalized substring kernel over a mixture of words and word classes to extract protein interactions from biomedical corpora and semantic relations from newswire corpora. Particularly, they achieve the F1-score of 54.2 in extracting protein interactions from the AIMed corpus. Erkan et al. (2007) first define two similarity functions based on co-sine similarity and edit distance among dependency paths between two entities, and then incorporate them in semi-supervised learning for PPI extraction using SVM and KNN classifiers. Saetre et al. (2007) use a tree kernel over dependency structures from two parsers and achieve the F1-score of 52.0 for PPI extraction from the AIMed corpus. Airola et al. (2008) introduce an all-dependency-paths graph kernel to capture complex dependency relationships between words and attain a significant performance boost at the expense of computational complexity. They achieve the F1-score of 56.4 in PPI extraction from the AIMed corpus. Kim et al. (2010) adopt a walk-weighted subsequence kernel based on shortest dependency paths to explore various substructures such as e-walks, partial match, and non-contiguous paths. They achieve the F1-score of 56.7 on the AIMed corpus.",
	"cite_spans": [
	{
	"start": 107,
	"end": 129,
	"text": "Bunescu et al. (2005a)",
	"ref_id": "BIBREF2"
	},
	{
	"start": 416,
	"end": 435,
	"text": "Erkan et al. (2007)",
	"ref_id": "BIBREF10"
	},
	{
	"start": 669,
	"end": 689,
	"text": "Saetre et al. (2007)",
	"ref_id": "BIBREF28"
	},
	{
	"start": 827,
	"end": 847,
	"text": "Airola et al. (2008)",
	"ref_id": "BIBREF1"
	},
	{
	"start": 1115,
	"end": 1132,
	"text": "Kim et al. (2010)",
	"ref_id": "BIBREF15"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "PPI extraction in biomedical domain",
	"sec_num": null
	},
	{
	"text": "In the literature, various kernels-based methods (Zelenko et al., 2003; Culotta and Sorensen, 2004; Bunescu et al., 2005b; Zhang et al., 2006; Zhou et al., 2007a; Qian et al., 2008) have been widely used in semantic relation extraction in the newswire domain. In particular, Zhang et al. (2006) , Zhou et al. (2007a) and Qian et al. (2008) adopt convolution tree kernels (Collins and Duffy, 2001 ) over constituent parse trees and show great success with comparable or even better performance than feature-based ones, motivated by the pioneer work of Moschitti et al. (2004; on semantic role labeling.",
	"cite_spans": [
	{
	"start": 49,
	"end": 71,
	"text": "(Zelenko et al., 2003;",
	"ref_id": "BIBREF30"
	},
	{
	"start": 72,
	"end": 99,
	"text": "Culotta and Sorensen, 2004;",
	"ref_id": "BIBREF7"
	},
	{
	"start": 100,
	"end": 122,
	"text": "Bunescu et al., 2005b;",
	"ref_id": "BIBREF3"
	},
	{
	"start": 123,
	"end": 142,
	"text": "Zhang et al., 2006;",
	"ref_id": "BIBREF31"
	},
	{
	"start": 143,
	"end": 162,
	"text": "Zhou et al., 2007a;",
	"ref_id": "BIBREF35"
	},
	{
	"start": 163,
	"end": 181,
	"text": "Qian et al., 2008)",
	"ref_id": "BIBREF26"
	},
	{
	"start": 275,
	"end": 294,
	"text": "Zhang et al. (2006)",
	"ref_id": "BIBREF31"
	},
	{
	"start": 297,
	"end": 316,
	"text": "Zhou et al. (2007a)",
	"ref_id": "BIBREF35"
	},
	{
	"start": 321,
	"end": 339,
	"text": "Qian et al. (2008)",
	"ref_id": "BIBREF26"
	},
	{
	"start": 371,
	"end": 395,
	"text": "(Collins and Duffy, 2001",
	"ref_id": "BIBREF6"
	},
	{
	"start": 551,
	"end": 574,
	"text": "Moschitti et al. (2004;",
	"ref_id": "BIBREF21"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Semantic relation extraction in newswire domain",
	"sec_num": null
	},
	{
	"text": "While convolution kernels (Haussler et al., 1999) can effectively capture structured information in discrete objects, the key problem for tree kernel-based methods lies largely in how to appropriately represent structured syntactic information inherent in relation instances. Zhang et al. (2006) discover that the Shortest Path-enclosed Tree (SPT) achieves the best performance among five tree setups. Zhou et al. (2007a) further extend it to Context-Sensitive Shortest Pathenclosed Tree (CS-SPT), which includes necessary predicate-linked path information. Qian et al. (2008) propose to automatically determine the appropriate part of a constituent parse tree by considering constituent dependencies on each node along the shortest path between two entity mentions and discarding irrelevant nodes. However, their adopted constituent dependency rules are manually constructed and thus difficult to adapt to other domains and languages. Figure 1 . Different tree setups for a PPI instance between PROT1 and PROT2 from sentence \"Association between PROT1 and cyclin B1 / PROT2 was detected in the HeLa cells.\" in the AIMed corpus",
	"cite_spans": [
	{
	"start": 26,
	"end": 49,
	"text": "(Haussler et al., 1999)",
	"ref_id": "BIBREF13"
	},
	{
	"start": 276,
	"end": 295,
	"text": "Zhang et al. (2006)",
	"ref_id": "BIBREF31"
	},
	{
	"start": 402,
	"end": 421,
	"text": "Zhou et al. (2007a)",
	"ref_id": "BIBREF35"
	},
	{
	"start": 558,
	"end": 576,
	"text": "Qian et al. (2008)",
	"ref_id": "BIBREF26"
	}
	],
	"ref_spans": [
	{
	"start": 936,
	"end": 944,
	"text": "Figure 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Semantic relation extraction in newswire domain",
	"sec_num": null
	},
	{
	"text": "In this paper, we make use of the shortest dependency path in the dependency tree to refine the constituent parse tree. Specifically, all the words which appear on the shortest dependency path, together with their associated constituents, are kept in the constituent parse tree, while other constituents are removed, forming a Shortest Dependency Path-directed Constituent Parse Tree (SDP-CPT).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Semantic relation extraction in newswire domain",
	"sec_num": null
	},
	{
	"text": "This section first illustrates the limitations of commonly-used constituent parse tree setups, then emphasizes the importance of shortest dependency path in representing the constituent parse tree, and finally presents the shortest dependency path-directed constituent parse tree (SDP-CPT).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Constituent Structure Representation",
	"sec_num": "3"
	},
	{
	"text": "It is widely acknowledged that the key problem for the success of tree kernel-based semantic relation extraction is how to represent the constituent parse tree in a precise and concise manner. Zhang et al. (2006) explore five kinds of tree setups and find that the Shortest Path-enclosed Tree (SPT) achieves the best performance. However, unlike the locality of semantic relations in the newswire domain (Zhou et al., 2005) , most of PPI instances in the biomedical domain spans a relatively long distance, leading to more complexity and diversity (Bunescu et al., 2005c; Airola et al., 2008) . Therefore, it is not surprising that previous tree kernels over constituent parse trees have not yet achieved promising results for PPI extraction just as they do in the news domain. Miyao et al. (2008) conduct a comprehensive comparison of different syntactic representations for PPI extraction and find that the phrase structure tree in the form of the constituent parse tree (called PTB in their paper) performs significantly worse than the other representations. Tikk et al. (2010) extensively compares different kernelbased methods on PPI extraction and show that the tree kernel over the constituent parse tree only achieves the F1-score of 34.6 on the AIMed corpus. Actually, our preliminary experiment on PPI extraction via the convolution tree kernel over SPT only achieves the F1-score of about 47 on the AIMed corpus. Such poor performance can be justified to a certain extent via a typical instance as illustrated in Figure 1 , where the interaction between PROT1 and PROT2 (their actual names have been replaced) can be only determined by the overall constituent structure of the sentence. Obviously, SPT will fail to identify this interaction instance since SPT ignores the constituents outside the shortest path ( Figure 1 :",
	"cite_spans": [
	{
	"start": 193,
	"end": 212,
	"text": "Zhang et al. (2006)",
	"ref_id": "BIBREF31"
	},
	{
	"start": 404,
	"end": 423,
	"text": "(Zhou et al., 2005)",
	"ref_id": "BIBREF34"
	},
	{
	"start": 548,
	"end": 571,
	"text": "(Bunescu et al., 2005c;",
	"ref_id": "BIBREF4"
	},
	{
	"start": 572,
	"end": 592,
	"text": "Airola et al., 2008)",
	"ref_id": "BIBREF1"
	},
	{
	"start": 778,
	"end": 797,
	"text": "Miyao et al. (2008)",
	"ref_id": "BIBREF20"
	},
	{
	"start": 1062,
	"end": 1080,
	"text": "Tikk et al. (2010)",
	"ref_id": "BIBREF29"
	}
	],
	"ref_spans": [
	{
	"start": 1524,
	"end": 1532,
	"text": "Figure 1",
	"ref_id": null
	},
	{
	"start": 1824,
	"end": 1832,
	"text": "Figure 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Limitations of Current Tree Setups",
	"sec_num": "3.1"
	},
	{
	"text": "T 2 : SPT).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Limitations of Current Tree Setups",
	"sec_num": "3.1"
	},
	{
	"text": "For the Context-Sensitive SPT (CS-SPT), as proposed in Zhou et al. (2007a) , which extends necessary predicate-linked path information outside SPT, some critical information is still missing while there exists some noisy information. For the instance as shown in Figure 1 (T 3 : CS-SPT), although the word \"detected\" and its asso-ciated constituents are added, the more important portion of \"association between\" and their associated constituents are still missing while the noisy words \"cyclin B1 / \" still remaining.",
	"cite_spans": [
	{
	"start": 55,
	"end": 74,
	"text": "Zhou et al. (2007a)",
	"ref_id": "BIBREF35"
	}
	],
	"ref_spans": [
	{
	"start": 263,
	"end": 271,
	"text": "Figure 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Limitations of Current Tree Setups",
	"sec_num": "3.1"
	},
	{
	"text": "In order to overcome the shortcomings in SPT and CS-SPT, Qian et al. (2008) propose a dynamic syntactic parse tree (DSPT) by exploiting constituent dependencies to refine the constituent parse tree. Specifically, they manually devise five categories of constituent dependencies, motivated by various kinds of lexical dependencies. When refining each node along the shortest path in the constituent parse tree, these constituent dependencies are used to determine how to remove or reduce futile constituents, eventually leading to a more precise and concise parse tree structure. However, this tree structure still suffers from the following three shortcomings: 1) It disregards the constituents beyond the lowest common ancestor to the tree root, similar to CS-SPT as proposed in Zhou et al. (2007a) . This may be largely due to the locality of semantic relations as defined in the ACE RDC corpus, which Zhou et al (2007a) and Qian et al. (2008) tackle.",
	"cite_spans": [
	{
	"start": 57,
	"end": 75,
	"text": "Qian et al. (2008)",
	"ref_id": "BIBREF26"
	},
	{
	"start": 780,
	"end": 799,
	"text": "Zhou et al. (2007a)",
	"ref_id": "BIBREF35"
	},
	{
	"start": 904,
	"end": 922,
	"text": "Zhou et al (2007a)",
	"ref_id": "BIBREF35"
	},
	{
	"start": 927,
	"end": 945,
	"text": "Qian et al. (2008)",
	"ref_id": "BIBREF26"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Limitations of Current Tree Setups",
	"sec_num": "3.1"
	},
	{
	"text": "2) The rules adopted to tackle constituent dependencies are manually constructed and thus may not be easily adapted to other domains and languages. For example, while the constituent dependencies related to noun phrases are effective in the newswire domain (e.g. the ACE RDC corpus), this may not be true for PPI extraction in the biomedical literature. 3) The constituent dependencies have been divided into only five categories. Such division may be too coarse to reflect the substantial difference between various kinds of dependencies (considering there are 55 kinds of minor-typed dependencies for the Stanford Dependency representation).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Limitations of Current Tree Setups",
	"sec_num": "3.1"
	},
	{
	"text": "In this paper, we attempt to address these problems by considering the shortest dependency path in the dependency parse tree for reshaping the constituent parse tree in a principled way in the context of PPI extraction from the biomedical literature.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Limitations of Current Tree Setups",
	"sec_num": "3.1"
	},
	{
	"text": "Lexical dependencies can indicate both local and long-range relationships among words occurring in the same sentence. Such dependency relationships offer a condensed representation of the information necessary to assess the relationship between two proteins or entities. In order to capture the necessary information inherent in the depedency parse tree for extracting PPI instances, various kernels based on dependency paths, such as edit distance kernel (Erkan et al., 2007) , alldependency-path graph kernel (Airola et al., 2008) , and walk-weighted subsequence kernels (Kim et al., 2010) have been proposed. Likewise for semantic relation extraction in the newswire domain, the kernels on dependency trees (Culotta and Sorensen, 2004) and the shortest dependency path (Bunescu et al., 2005b) have been proposed. One common characteristic to these kernels is that they all contain the shortest dependency path and usually assign more weights to them than to other ones, similar to the graph kernel proposed by Airola et al. (2008) . This indicates the importance of the shortest dependency path over other paths in the dependency path tree or the dependency graph.",
	"cite_spans": [
	{
	"start": 456,
	"end": 476,
	"text": "(Erkan et al., 2007)",
	"ref_id": "BIBREF10"
	},
	{
	"start": 511,
	"end": 532,
	"text": "(Airola et al., 2008)",
	"ref_id": "BIBREF1"
	},
	{
	"start": 573,
	"end": 591,
	"text": "(Kim et al., 2010)",
	"ref_id": "BIBREF15"
	},
	{
	"start": 710,
	"end": 738,
	"text": "(Culotta and Sorensen, 2004)",
	"ref_id": "BIBREF7"
	},
	{
	"start": 772,
	"end": 795,
	"text": "(Bunescu et al., 2005b)",
	"ref_id": "BIBREF3"
	},
	{
	"start": 1013,
	"end": 1033,
	"text": "Airola et al. (2008)",
	"ref_id": "BIBREF1"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Shortest Dependency Path",
	"sec_num": "3.2"
	},
	{
	"text": "Currently, there are two established dependency representations available, viz. CoNLL scheme (adopted by CoNLL'2007 and CoNLL'2008 Shared tasks) (Nivre et al., 2007; Surdeanu et al., 2008) and Standford scheme (adopted by Stanford parser) (de Marneffe et al., 2006) . These two schemes differ significantly in the representation of passive construction, position of auxiliary and modal verb, or coordination. It is generally acknowledged that the Stanford scheme is closer to the targeted semantic representation from the perspective of relation extraction (Buyko and Hahn, 2010) . Particularly, among the four styles of Stanford representations, \"collapsed dependency\" can much simplify patterns in relation extraction since dependencies involving preposition, conjunct as well as referent of relative clause are effectively collapsed to reflect direct dependencies between content words. Therefore, the collapsed variant of Stanford scheme is adopted in this paper to refine the constituent parse tree as described in the next subsection.",
	"cite_spans": [
	{
	"start": 105,
	"end": 119,
	"text": "CoNLL'2007 and",
	"ref_id": null
	},
	{
	"start": 120,
	"end": 144,
	"text": "CoNLL'2008 Shared tasks)",
	"ref_id": null
	},
	{
	"start": 145,
	"end": 165,
	"text": "(Nivre et al., 2007;",
	"ref_id": "BIBREF24"
	},
	{
	"start": 166,
	"end": 188,
	"text": "Surdeanu et al., 2008)",
	"ref_id": "BIBREF27"
	},
	{
	"start": 239,
	"end": 265,
	"text": "(de Marneffe et al., 2006)",
	"ref_id": "BIBREF8"
	},
	{
	"start": 557,
	"end": 579,
	"text": "(Buyko and Hahn, 2010)",
	"ref_id": "BIBREF5"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Shortest Dependency Path",
	"sec_num": "3.2"
	},
	{
	"text": "Considering the importance of dependency path in PPI extraction and the effectiveness of employing dependency information to refine the constituent parse tree for tree kernel-based semantic relation extraction in the newswire domain, it is a natural idea to automatically generate the proper constituent parse tree with the help of the shortest dependency path. Specifically, we can reshape the constituent parse tree by making use of the shortest dependency path between two proteins. Figure 2 describes the procedure to generate the Shortest Dependency Path-directed Constituent Parse Tree (SDP-CPT).",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 486,
	"end": 494,
	"text": "Figure 2",
	"ref_id": "FIGREF1"
	}
	],
	"eq_spans": [],
	"section": "SDP-CPT: Shortest Dependency Pathdirected Constituent Parse Tree",
	"sec_num": "3.3"
	},
	{
	"text": "Step 3(a) in Figure 2 is necessary since a dependency tuple of the type \"prep_xx(governor, dependent)\" implies a relationship between the preposition xx and the dependent, which is important to PPI. For Step 3(b), when a word on which the two proteins are directly or indirectly dependent is discovered, it is natural to add this path for maintaining the integrality of SDP-CPT.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 13,
	"end": 21,
	"text": "Figure 2",
	"ref_id": "FIGREF1"
	}
	],
	"eq_spans": [],
	"section": "Note that",
	"sec_num": null
	},
	{
	"text": "Input: a sentence and two proteins in it Output: an SDP-CPT Steps: 1) Given the input sentence, generate the constituent parse tree using a constituent parser, and various dependency tuples using a dependency parser. 2) Given the two proteins, extract the shortest constituent path (SCP, i.e. the shortest path-enclosed tree) from the constituent parse tree and construct the shortest dependency path (SDP) from the dependency tuples. 3) For each word along the SDP, add the corresponding leaf word node and its upper constituents to the SCP. Particularly, a) when the dependency type is \"prep_xx\", such as \"prep_of\", the preposition xx and its associated constituent are also added; b) when the word to be added is outside the SCP, a new path from the current lowest common ancestor to one of the added words' ancestors is also added. 4) Merge any two consecutive NP/VP nodes along the paths into a single one. In order to demonstrate the process of generating a SDP-CPT, we take the sentence and the two proteins shown in Figure 1 as an example. Figure 3 illustrates the detailed generation process. First, the shortest dependency path (SDP) and the shortest constituent path (SCP) are generated as depicted in Figure 3 (a) and 3(b) respectively. Then, every word in the SDP is added into the SCP together with its associated constituents.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 1024,
	"end": 1032,
	"text": "Figure 1",
	"ref_id": null
	},
	{
	"start": 1048,
	"end": 1056,
	"text": "Figure 3",
	"ref_id": null
	},
	{
	"start": 1213,
	"end": 1221,
	"text": "Figure 3",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Note that",
	"sec_num": null
	},
	{
	"text": "In this case, since the two protein names in the SDP share a common ancestor \"Association\", the word \"Association\" together with its constituent ancestors are added into the SCP and a new path \"NP PP NP\" is created as rendered by the dashed lines. Finally, since the dependency type between \"PROT1\" and \"Association\" is prep_between, the preposition word \"between\" and its constituent ancestors are added into the SCP as rendered by the dotted lines. Since no further post-processing is necessary in this example, SDP-CPT is eventually formed. Compared to other tree setups in Figure 1 , namely SPT, CS-SPT and DSPT, obviously SDP-CPT is much more concise and precise for this PPI instance.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 577,
	"end": 585,
	"text": "Figure 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Note that",
	"sec_num": null
	},
	{
	"text": "This section systematically evaluates the performance of our shortest dependency pathdirected constituent parse tree (SDP-SPT) on PPI extraction across several major PPI corpora.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Experimentation",
	"sec_num": "4"
	},
	{
	"text": "In order to fairly compare our work with other PPI extraction systems, we use five PPI corpora, i.e., AIMed (Bunescu et al., 2005a) , BioInfer (Pyysalo et al., 2007) , HPRD50 (Fundel et al., 2007) , IEPA (Ding et al., 2002) and LLL (N\u00e9dellec, 2005) . Particularly, most of the evaluation is done on the widely-used AIMed corpus, which contains 177 Medline abstracts with PPI instances, and 48 abstracts without any PPI instances. Totally, there are 4,084 protein references and around 1,000 annotated protein-protein interactions in this data set.",
	"cite_spans": [
	{
	"start": 108,
	"end": 131,
	"text": "(Bunescu et al., 2005a)",
	"ref_id": "BIBREF2"
	},
	{
	"start": 143,
	"end": 165,
	"text": "(Pyysalo et al., 2007)",
	"ref_id": "BIBREF25"
	},
	{
	"start": 175,
	"end": 196,
	"text": "(Fundel et al., 2007)",
	"ref_id": "BIBREF11"
	},
	{
	"start": 204,
	"end": 223,
	"text": "(Ding et al., 2002)",
	"ref_id": "BIBREF9"
	},
	{
	"start": 232,
	"end": 248,
	"text": "(N\u00e9dellec, 2005)",
	"ref_id": "BIBREF23"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Data Sets and Preprocessing",
	"sec_num": "4.1"
	},
	{
	"text": "In this paper, a potential PPI instance is generated for any pair of two proteins in a sentence That is, if a sentence contains n proteins, ( n 2 ) protein pairs are generated. In particular, all the self-interactions (59 instances) are removed and all the PPI instances with nested protein names are retained (154 instances), as adopted in most literature. Eventually, 1000 positive instances and 4834 negative instances are generated. Besides, for a potential PPI instance, the two involved proteins are replaced by PROT1 and PROT2 respectively in order to blind the learner for fair comparison with other work. Finally, all the sentences in these corpora are parsed using the Stanford Parser 1 to generate both the con-stituent parse trees and their corresponding dependency tuples.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Data Sets and Preprocessing",
	"sec_num": "4.1"
	},
	{
	"text": "In our experimentation, we select Support Vector Machines (SVM) as the classifier since SVM represents the state-of-the-art in the machine learning research community. In particular, we use the SVM light (Joachims, 1998) with the convolution tree kernel function SVM light -TK (Moschitti, 2004) 2 to compute the similarity between two constituent parse trees.",
	"cite_spans": [
	{
	"start": 204,
	"end": 220,
	"text": "(Joachims, 1998)",
	"ref_id": "BIBREF14"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Classifier and Evaluation Metrics",
	"sec_num": "4.2"
	},
	{
	"text": "Evaluation is done using 10-fold documentlevel cross-validation, each of which contains 90% of documents as the training data and 10% as the test data. Particularly, for the AIMed corpus we apply the exactly same 10-fold split as widely used in a series of relevant studies (e.g., Bunescu et al., 2005a; Giuliano et al., 2006) . Following conventions, the parameters C for SVM is set to the ratio of negative instances to positive ones in respective corpora, and \u03bb for the convolution tree kernel is set to default 0.4. Furthermore, the OAOD (One Answer per Occurrence in the Document) strategy is adopted, which means that the correct interaction must be extracted for each occurrence. This guarantees the maximal use of the available data, and more importantly, allows fair comparison with relevant work. All the experiments are evaluated using commonly-used Precision (P), Recall (R) and harmonic F1-score (F1). As an alternative to F1-score, the AUC (area under the receiver operating characteristics curve) score is proved to be invariant to the class distribution of the test dataset. Therefore, we also provide the AUC score of our system for referrence as by Airola et al. (2008) . Table 1 reports the performance of PPI extraction on the AIMed corpus corresponding to different lengths of the dependency paths using all kinds of dependency types. Here, two partial dependency paths on SDP, starting from each of the two proteins respectively, are utilized to generate the tree representation. The length of these two paths is shown in the 1 st column. The words corresponding to the nodes on these two paths together with these words' associated constituents in the parse tree are added to SCP. For ex-ample, the length of 0 (L0) means that not a single word or constituent will be added to SCP, while the length of 1 (L1) means that the words corresponding to the parents of two proteins on SDP and these words' associated constituents in the parse tree are added to SCP. Meanwhile, the performance of the SPT setup is also listed as the baseline for comparison. This table shows that the constituent parse tree directed by the shortest dependency path (SDP-CPT) achieves the best performance of 59.6/54.3/56.7/82.7 in P/R/F1/AUC, significantly outperforming SPT by 9.6 units in F1 and 2.8 units in AUC largely due to the substantial increase in recall. This indicates that SDP-CPT can remove much noise in SPT while adding some useful information. It also shows",
	"cite_spans": [
	{
	"start": 281,
	"end": 303,
	"text": "Bunescu et al., 2005a;",
	"ref_id": "BIBREF2"
	},
	{
	"start": 304,
	"end": 326,
	"text": "Giuliano et al., 2006)",
	"ref_id": "BIBREF12"
	},
	{
	"start": 1167,
	"end": 1187,
	"text": "Airola et al. (2008)",
	"ref_id": "BIBREF1"
	}
	],
	"ref_spans": [
	{
	"start": 1190,
	"end": 1197,
	"text": "Table 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Classifier and Evaluation Metrics",
	"sec_num": "4.2"
	},
	{
	"text": "The performance of the SCP corresponding to the length of 0 is lowest, since they contain no information derived from the shortest dependency path. With the increase of the length of dependency paths, more and more useful information derived from SDP is included in the constituent parse tree and the performance reaches the highest for SDP-CPT (all the words corresponding to all the nodes on the SDP and their associated constituents are added). In summary, the above results suggest that SDP-CPT can achieve the best performance. Therefore, all the subsequent experiments adopt the SDP-CPT setup unless specified. Table 2 compares the contribution of various kinds of dependencies in SDP-CPT on the AIMed corpus. All the typed dependency relations are grouped into 4 major classes, namely Modifier, Argument, Conjunction and Others. For every major type, minor dependency types, if any exists, are further ordered by their potential importance. The percentage of occurring frequency with which each minor type is employed when generating the SDP-CPT with respect to the total number of dependency tuples is listed in Column 2. Particularly, the tree setup without using any dependency type, which corresponds to that with the length of 0 (SCP) in Table 1 , is displayed at the top row. Furthermore, the dependency types are added in two different ways:",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 617,
	"end": 624,
	"text": "Table 2",
	"ref_id": null
	},
	{
	"start": 1250,
	"end": 1257,
	"text": "Table 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Comparison of different lengths of dependency paths on the AIMed corpus",
	"sec_num": null
	},
	{
	"text": "Individual: the dependency types are added individually with their performance scores shown inside the parentheses; Accumulative: the dependency types are incrementally added one by one with their performance scores shown outside the parentheses. The \"+\" sign before the type means that its addition can boost the performance in F1score or AUC score and thus will be passed down to the next iteration. Table 2 shows that with the addition of all Argument types and all Modifier types, the SDP-CPT attains the best performance of 59.1/57.6/58.1/83.3 in P/R/F1/AUC as shown in bold fonts, outperforming the SDP-CPT with all dependency types added (59.6/54.3/56.7/82.7 in P/R/F1/AUC). Particularly, it shows",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 402,
	"end": 409,
	"text": "Table 2",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Contribution of different kinds of dependencies on the AIMed corpus",
	"sec_num": null
	},
	{
	"text": "The dependency types of subj, obj, prep and nn yield substantial performance improve-ment both in the accumulative mode and in the individual mode;",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Contribution of different kinds of dependencies on the AIMed corpus",
	"sec_num": null
	},
	{
	"text": "The dependency types of Conjunction and Others harm the performance in the accumulative mode, though Conjunction improves the performance in the individual mode; It is interesting to note that while the dependency types of arg-others and mod-others harm the performance in the individual mode, they slightly improve the performance in the accumulative mode.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Contribution of different kinds of dependencies on the AIMed corpus",
	"sec_num": null
	},
	{
	"text": "Since the governors of subj and obj types are verbs, those of the prep type are nouns and prepositions, and those of the nn type are nouns, the above results are consistent with our observation that some verbs like \"bind\" or \"interact\", some prepositions like \"with\" or \"of\", and some nouns like \"interaction\" or \"expression\", on which two proteins are directly or indirectly dependent, are particularly important for PPI extraction. Henceforth, in the following experiments all the Argument and Modifier types are included while the Conjunction and Others types are excluded. Table 3 compares the performance of F1-score (outside parenthesis) and AUC (inside parentheses) between SDP-CPT and the previously-used tree setups across major PPI corpora. Particularly, SPT is used as a baseline and for comparison we re-implement two other effective tree setups for semantic relation extraction in the newswire domain, i.e. CS-SPT (Zhou et al., 2007a) and DSPT (Qian et al., 2008) . Significance tests are conducted between each of them and the baseline.",
	"cite_spans": [
	{
	"start": 927,
	"end": 947,
	"text": "(Zhou et al., 2007a)",
	"ref_id": "BIBREF35"
	},
	{
	"start": 957,
	"end": 976,
	"text": "(Qian et al., 2008)",
	"ref_id": "BIBREF26"
	}
	],
	"ref_spans": [
	{
	"start": 577,
	"end": 584,
	"text": "Table 3",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Contribution of different kinds of dependencies on the AIMed corpus",
	"sec_num": null
	},
	{
	"text": "Additionally, the numbers of positive and negative instances in each corpus are reported in the 1st row and the performance scores of MCT (Minimum Complete Tree, the complete sub-tree rooted by the lowest common ancestor of the two proteins under consideration) are also listed in the 2 nd row for reference. The table shows Among all tree setups SDP-CPT performs best and significantly outperforms SPT consistently on most PPI corpora. CS-SPT slightly outperforms SPT on most corpora while DSPT performs divergently on different corpora. The reason that DSPT performs excellently in the newswire domain (Qian et al., 2008) but not so much for PPI extraction may be that the heuristic rules they use to prune the constituent trees are more suitable for the newswire domain, thus limiting their capability of domain adaptation. In summary, the above results suggest the superiority and generality of our SDP-CPT on various kinds of PPI corpora from the biomedical literature.",
	"cite_spans": [
	{
	"start": 604,
	"end": 623,
	"text": "(Qian et al., 2008)",
	"ref_id": "BIBREF26"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Contribution of different kinds of dependencies on the AIMed corpus",
	"sec_num": null
	},
	{
	"text": "Our SDP-CPT kernel 59.1 57.6 58.1 Table 4 . Performance comparison of kernel-based PPI extraction systems on the AIMed corpus Table 4 compares our kernel-based system with other state-of-the-art kernel-based ones on the AIMed corpus using the exactly same 10-fold data splitting. It shows that our individual kernelbased system performs better than all the other individual kernel-based systems on the AIMed corpus. Particularly, our SDP-CPT kernel significantly outperforms the Partial Tree kernel over constituent parse trees (Tikk et al., 2010) . It even significantly outperforms the composite kernel combining BOW and constituent parse trees (Miyao et al., 2008) . Although our individual kernel performs worse than the composite kernels as adopted by Miyao et al. (2008) and Miwa et al. (2009) , the strength of our kernelbased system lies in the simplicity of our shortest dependency path-directed constituent parse tree.",
	"cite_spans": [
	{
	"start": 528,
	"end": 547,
	"text": "(Tikk et al., 2010)",
	"ref_id": "BIBREF29"
	},
	{
	"start": 647,
	"end": 667,
	"text": "(Miyao et al., 2008)",
	"ref_id": "BIBREF20"
	},
	{
	"start": 757,
	"end": 776,
	"text": "Miyao et al. (2008)",
	"ref_id": "BIBREF20"
	},
	{
	"start": 781,
	"end": 799,
	"text": "Miwa et al. (2009)",
	"ref_id": "BIBREF19"
	}
	],
	"ref_spans": [
	{
	"start": 34,
	"end": 41,
	"text": "Table 4",
	"ref_id": null
	},
	{
	"start": 126,
	"end": 133,
	"text": "Table 4",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "PPI extraction systems P(%) R(%) F1",
	"sec_num": null
	},
	{
	"text": "This paper presents a principled way to automatically generate the constituent parse tree for PPI extraction by making use of the shortest dependency path between two proteins. Although previous research indicates the difficulty of employing constituent parse tree information for PPI extraction due to the relatively long distance between two proteins, our detailed analysis and evaluation indicate that the constituent parse tree can achieve promising results for PPI extraction. Moreover, our dependency-directed constituent parse tree structure provides a general way to automatically determine the constituent parse tree for a wide class of related learning tasks, such as semantic relation extraction, semantic role labeling and even coreference resolution. For future work, we would like to apply our approach to other NLP tasks. Meanwhile, we will investigate the effect of constituent parse information on dependency-based relational learning in better exploring the synergy between dependency and constituent-based syntactic information.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conclusion and Future Work",
	"sec_num": "5"
	},
	{
	"text": "http://nlp.stanford.edu/software/lex-parser.shtml",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "http://ai-nlp.info.uniroma2.it/moschitti/",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	}
	],
	"back_matter": [
	{
	"text": "This research is supported by Projects 60873150, 60970056, and 90920004 under the National Natural Science Foundation of China, and the Project BK2010219 under the Provincial Natural Science Foundation of Jiangsu, China.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Acknowledgement",
	"sec_num": null
	}
	],
	"bib_entries": {
	"BIBREF0": {
	"ref_id": "b0",
	"title": "The Automatic Content Extraction (ACE) Projects",
	"authors": [],
	"year": 2002,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "ACE. 2002-2007. The Automatic Content Extraction (ACE) Projects. 2007. http//www.ldc.upenn.edu/ Projects/ACE/.",
	"links": null
	},
	"BIBREF1": {
	"ref_id": "b1",
	"title": "All-paths graph kernel for protein-protein interaction extraction with evaluation of cross corpus learning",
	"authors": [
	{
	"first": "A",
	"middle": [],
	"last": "Airola",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [],
	"last": "Pyysalo",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Bj\u00f6rne",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Pahikkala",
	"suffix": ""
	},
	{
	"first": "F",
	"middle": [],
	"last": "Ginter",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Salakoski",
	"suffix": ""
	}
	],
	"year": 2008,
	"venue": "BMC Bioinformatics",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Airola A., Pyysalo S., Bj\u00f6rne J., Pahikkala T., Ginter F. and Salakoski T. 2008. All-paths graph kernel for protein-protein interaction extraction with evalua- tion of cross corpus learning. BMC Bioinformatics.",
	"links": null
	},
	"BIBREF2": {
	"ref_id": "b2",
	"title": "Subsequence kernels for relation extraction",
	"authors": [
	{
	"first": "R",
	"middle": [],
	"last": "Bunescu",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Mooney",
	"suffix": ""
	}
	],
	"year": 2005,
	"venue": "Proceedings of NIPS'05",
	"volume": "",
	"issue": "",
	"pages": "171--178",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Bunescu R. and Mooney R. 2005a. Subsequence ker- nels for relation extraction. In Proceedings of NIPS'05: 171-178. December 2005.",
	"links": null
	},
	"BIBREF3": {
	"ref_id": "b3",
	"title": "A shortest path dependency kernel for relation extraction",
	"authors": [
	{
	"first": "R",
	"middle": [],
	"last": "Bunescu",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Mooney",
	"suffix": ""
	}
	],
	"year": 2005,
	"venue": "Proceedings of EMNIP'05",
	"volume": "",
	"issue": "",
	"pages": "724--731",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Bunescu R. and Mooney R.. 2005b. A shortest path dependency kernel for relation extraction. In Pro- ceedings of EMNIP'05: 724-731. Vancover, B.C.",
	"links": null
	},
	"BIBREF4": {
	"ref_id": "b4",
	"title": "Comparative Experiments on learning information extractors for Proteins and their interactions",
	"authors": [
	{
	"first": "R",
	"middle": [],
	"last": "Bunescu",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Ge",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Kate",
	"suffix": ""
	},
	{
	"first": "E",
	"middle": [],
	"last": "Marcotte",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Mooney",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Ramani",
	"suffix": ""
	},
	{
	"first": "Y",
	"middle": [],
	"last": "Wong",
	"suffix": ""
	}
	],
	"year": 2005,
	"venue": "Journal of Artificial Intelligence in Medicine",
	"volume": "33",
	"issue": "2",
	"pages": "139--155",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Bunescu R., Ge R., Kate R., Marcotte E., Mooney R., Ramani A. and Wong Y. 2005c. Comparative Ex- periments on learning information extractors for Proteins and their interactions. Journal of Artificial Intelligence in Medicine, 33(2): 139-155.",
	"links": null
	},
	"BIBREF5": {
	"ref_id": "b5",
	"title": "Evaluating the impact of alternative dependency graph encodings on solving event extraction tasks",
	"authors": [
	{
	"first": "E",
	"middle": [],
	"last": "Buyko",
	"suffix": ""
	},
	{
	"first": "U",
	"middle": [],
	"last": "Hahn",
	"suffix": ""
	}
	],
	"year": 2010,
	"venue": "Proceedings of EMNLP'2010",
	"volume": "",
	"issue": "",
	"pages": "982--992",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Buyko E. and Hahn U. 2010. Evaluating the impact of alternative dependency graph encodings on solving event extraction tasks. In Proceedings of EMNLP'2010: 982-992.",
	"links": null
	},
	"BIBREF6": {
	"ref_id": "b6",
	"title": "Convolution Kernels for Natural Language",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Collins",
	"suffix": ""
	},
	{
	"first": "N",
	"middle": [],
	"last": "Duffy",
	"suffix": ""
	}
	],
	"year": 2001,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "625--632",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Collins M. and Duffy N. 2001. Convolution Kernels for Natural Language. NIPS 2001: 625-632.",
	"links": null
	},
	"BIBREF7": {
	"ref_id": "b7",
	"title": "Dependency Tree Kernels for Relation Extraction",
	"authors": [
	{
	"first": "A",
	"middle": [],
	"last": "Culotta",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Sorensen",
	"suffix": ""
	}
	],
	"year": 2004,
	"venue": "Proceedings of ACL'04",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Culotta A. and Sorensen J. 2004. Dependency Tree Kernels for Relation Extraction. In Proceedings of ACL'04. July 2004. Barcelona, Spain.",
	"links": null
	},
	"BIBREF8": {
	"ref_id": "b8",
	"title": "Generating typed dependency parses from phrase structure parses",
	"authors": [
	{
	"first": "M.-C",
	"middle": [],
	"last": "De Marneffe",
	"suffix": ""
	},
	{
	"first": "B",
	"middle": [],
	"last": "Maccartney",
	"suffix": ""
	},
	{
	"first": "C",
	"middle": [
	"D"
	],
	"last": "Manning",
	"suffix": ""
	}
	],
	"year": 2006,
	"venue": "LREC",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "de Marneffe M.-C., MacCartney B. and Manning C. D. 2006. Generating typed dependency parses from phrase structure parses. In LREC 2006.",
	"links": null
	},
	"BIBREF9": {
	"ref_id": "b9",
	"title": "Mining medline: abstracts, sentences, or phrases?",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Ding",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [],
	"last": "Berleant",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [],
	"last": "Nettleton",
	"suffix": ""
	},
	{
	"first": "E",
	"middle": [],
	"last": "Wurtele",
	"suffix": ""
	}
	],
	"year": 2002,
	"venue": "Pacific Symposium on Biocomputing",
	"volume": "",
	"issue": "",
	"pages": "326--337",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Ding J., Berleant D., Nettleton D. and Wurtele E. 2002. Mining medline: abstracts, sentences, or phrases? Pacific Symposium on Biocomputing, pages 326-337.",
	"links": null
	},
	"BIBREF10": {
	"ref_id": "b10",
	"title": "Semi-Supervised Classification for Extracting Protein Interaction Sentences using Dependency Parsing",
	"authors": [
	{
	"first": "G",
	"middle": [],
	"last": "Erkan",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Ozgur",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [
	"R"
	],
	"last": "Radev",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "Proceedings of EMNLP-CoNLL'07",
	"volume": "",
	"issue": "",
	"pages": "228--237",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Erkan G., Ozgur A. and Radev D.R. 2007. Semi- Supervised Classification for Extracting Protein In- teraction Sentences using Dependency Parsing, In Proceedings of EMNLP-CoNLL'07: 228-237. June, 2007, Prague, Czech.",
	"links": null
	},
	"BIBREF11": {
	"ref_id": "b11",
	"title": "Relexrelation extraction using dependency parse trees",
	"authors": [
	{
	"first": "K",
	"middle": [],
	"last": "Fundel",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Kuffner",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Zimmer",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "Bioinformatics",
	"volume": "23",
	"issue": "3",
	"pages": "365--371",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Fundel K., Kuffner R. and Zimmer R. 2007. Relex- relation extraction using dependency parse trees. Bioinformatics, 23(3):365-371.",
	"links": null
	},
	"BIBREF12": {
	"ref_id": "b12",
	"title": "Exploiting Shallow Linguistic Information for Relation Extraction from Biomedical Literature",
	"authors": [
	{
	"first": "C",
	"middle": [],
	"last": "Giuliano",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Lavelli",
	"suffix": ""
	},
	{
	"first": "L",
	"middle": [],
	"last": "Romano",
	"suffix": ""
	}
	],
	"year": 2006,
	"venue": "Proceedings of EACL'06",
	"volume": "",
	"issue": "",
	"pages": "401--408",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Giuliano C., Lavelli A. and Romano L. 2006. Exploit- ing Shallow Linguistic Information for Relation Ex- traction from Biomedical Literature. In Proceedings of EACL'06: 401-408. April, 2006, Trento, Italy.",
	"links": null
	},
	"BIBREF13": {
	"ref_id": "b13",
	"title": "Convolution Kernels on Discrete Structures",
	"authors": [
	{
	"first": "D",
	"middle": [],
	"last": "Haussler",
	"suffix": ""
	}
	],
	"year": 1999,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Haussler D. 1999. Convolution Kernels on Discrete Structures. Technical Report UCS-CRL-99-10, University of California, Santa Cruz.",
	"links": null
	},
	"BIBREF14": {
	"ref_id": "b14",
	"title": "Text Categorization with Support Vector Machine: learning with many relevant features",
	"authors": [
	{
	"first": "T",
	"middle": [],
	"last": "Joachims",
	"suffix": ""
	}
	],
	"year": 1998,
	"venue": "Proceedings of European Conference on Machine Learning",
	"volume": "",
	"issue": "",
	"pages": "137--142",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Joachims T. 1998. Text Categorization with Support Vector Machine: learning with many relevant fea- tures. In Proceedings of European Conference on Machine Learning(ECML'1998): 137-142.",
	"links": null
	},
	"BIBREF15": {
	"ref_id": "b15",
	"title": "Walkweighted subsequence kernels for protein-protein interaction extraction",
	"authors": [
	{
	"first": "S",
	"middle": [],
	"last": "Kim",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Yoon",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Yang",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [],
	"last": "Park",
	"suffix": ""
	}
	],
	"year": 2010,
	"venue": "Journal of BMC Bioinformatics",
	"volume": "11",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Kim S., Yoon J., Yang J. and Park S. 2010. Walk- weighted subsequence kernels for protein-protein interaction extraction. Journal of BMC Bioinfor- matics, 2010, 11:107.",
	"links": null
	},
	"BIBREF16": {
	"ref_id": "b16",
	"title": "Kernel-Based Learning for Biomedical Relation extraction",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Li",
	"suffix": ""
	},
	{
	"first": "Z",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	},
	{
	"first": "X",
	"middle": [],
	"last": "Li",
	"suffix": ""
	},
	{
	"first": "H",
	"middle": [],
	"last": "Chen",
	"suffix": ""
	}
	],
	"year": 2008,
	"venue": "Journal of the American Society for Information Science and Technology",
	"volume": "59",
	"issue": "5",
	"pages": "756--769",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Li J., Zhang Z., Li X. and Chen H. 2008. Kernel- Based Learning for Biomedical Relation extraction. Journal of the American Society for Information Science and Technology, 59(5):756-769.",
	"links": null
	},
	"BIBREF17": {
	"ref_id": "b17",
	"title": "Dependency-Driven Feature-based Learning for Extracting Protein-Protein Interactions from Biomedical Text",
	"authors": [
	{
	"first": "B",
	"middle": [],
	"last": "Liu",
	"suffix": ""
	},
	{
	"first": "L",
	"middle": [
	"H"
	],
	"last": "Qian",
	"suffix": ""
	},
	{
	"first": "H",
	"middle": [
	"L"
	],
	"last": "Wang",
	"suffix": ""
	},
	{
	"first": "G",
	"middle": [
	"D"
	],
	"last": "Zhou",
	"suffix": ""
	}
	],
	"year": 2010,
	"venue": "Proceedings of COLING'2010",
	"volume": "",
	"issue": "",
	"pages": "757--765",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Liu B., Qian L.H., Wang H.L. and Zhou G.D. 2010. Dependency-Driven Feature-based Learning for Extracting Protein-Protein Interactions from Bio- medical Text. In Proceedings of COLING'2010: 757-765 (Poster).",
	"links": null
	},
	"BIBREF18": {
	"ref_id": "b18",
	"title": "Extracting protein-protein interaction information from biomedical text with SVM",
	"authors": [
	{
	"first": "T",
	"middle": [],
	"last": "Mitsumori",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Murata",
	"suffix": ""
	},
	{
	"first": "Y",
	"middle": [],
	"last": "Fukuda",
	"suffix": ""
	},
	{
	"first": "K",
	"middle": [],
	"last": "Doi",
	"suffix": ""
	},
	{
	"first": "H",
	"middle": [],
	"last": "Doi",
	"suffix": ""
	}
	],
	"year": 2006,
	"venue": "IEICE Transactions on Information and Systems",
	"volume": "",
	"issue": "8",
	"pages": "2464--2466",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Mitsumori T., Murata M., Fukuda Y., Doi K. and Doi H. 2006. Extracting protein-protein interaction in- formation from biomedical text with SVM. IEICE Transactions on Information and Systems, E89-D (8): 2464-2466.",
	"links": null
	},
	"BIBREF19": {
	"ref_id": "b19",
	"title": "A Rich Feature Vector for Protein-Protein Interaction Extraction from Multiple Corpora",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Miwa",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Saetre",
	"suffix": ""
	},
	{
	"first": "Y",
	"middle": [],
	"last": "Miyao",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Tsujii",
	"suffix": ""
	}
	],
	"year": 2009,
	"venue": "Proceedings of EMNLP'09",
	"volume": "",
	"issue": "",
	"pages": "121--130",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Miwa M., Saetre R., Miyao Y. and Tsujii J. 2009. A Rich Feature Vector for Protein-Protein Interaction Extraction from Multiple Corpora. In Proceedings of EMNLP'09: 121-130. August 2-7, 2009, Singa- pore.",
	"links": null
	},
	"BIBREF20": {
	"ref_id": "b20",
	"title": "Task-oriented evaluation of syntactic parsers and their representations",
	"authors": [
	{
	"first": "Y",
	"middle": [],
	"last": "Miyao",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Saetre",
	"suffix": ""
	},
	{
	"first": "K",
	"middle": [],
	"last": "Sagae",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Matsuzaki",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Tsujii",
	"suffix": ""
	}
	],
	"year": 2008,
	"venue": "Proceedings of the 45th Meeting of the Association for Computational Linguistics (ACL'08: HLT",
	"volume": "",
	"issue": "",
	"pages": "46--54",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Miyao Y., Saetre R., Sagae K., Matsuzaki T. and Tsu- jii J. 2008. Task-oriented evaluation of syntactic parsers and their representations. In Proceedings of the 45th Meeting of the Association for Computa- tional Linguistics (ACL'08: HLT): 46-54.",
	"links": null
	},
	"BIBREF21": {
	"ref_id": "b21",
	"title": "A Study on Convolution Kernels for Shallow Semantic Parsing",
	"authors": [
	{
	"first": "A",
	"middle": [],
	"last": "Moschitti",
	"suffix": ""
	}
	],
	"year": 2004,
	"venue": "Proceedings of ACL'2004",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Moschitti A. 2004. A Study on Convolution Kernels for Shallow Semantic Parsing. In Proceedings of ACL'2004. July 21-26, 2004. Barcelona, Spain.",
	"links": null
	},
	"BIBREF22": {
	"ref_id": "b22",
	"title": "Tree Kernels for Semantic Role Labeling, Special Issue on Semantic Role Labeling",
	"authors": [
	{
	"first": "A",
	"middle": [],
	"last": "Moschitti",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [],
	"last": "Pighin",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Basili",
	"suffix": ""
	}
	],
	"year": 2008,
	"venue": "Computational Linguistics",
	"volume": "34",
	"issue": "2",
	"pages": "194--224",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Moschitti A., Pighin D. and Basili R. 2008. Tree Ker- nels for Semantic Role Labeling, Special Issue on Semantic Role Labeling, Computational Linguistics. 34(2): 194-224.",
	"links": null
	},
	"BIBREF23": {
	"ref_id": "b23",
	"title": "Learning language in logic-genic interaction extraction challenge",
	"authors": [
	{
	"first": "C",
	"middle": [],
	"last": "N\u00e9dellec",
	"suffix": ""
	}
	],
	"year": 2005,
	"venue": "Proceedings of the LLL'05 Workshop",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "N\u00e9dellec C. 2005. Learning language in logic-genic interaction extraction challenge. In Proceedings of the LLL'05 Workshop.",
	"links": null
	},
	"BIBREF24": {
	"ref_id": "b24",
	"title": "The CoNLL",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Nivre",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Hall",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [],
	"last": "Kubler",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Mcdonald",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Nilsson",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [],
	"last": "Riedel",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [],
	"last": "Yuret",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "Shared Task on Dependency Parsing. EMNLP-CoNLL",
	"volume": "",
	"issue": "",
	"pages": "915--932",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Nivre J., Hall J., Kubler S., McDonald R., Nilsson J., Riedel S. and Yuret D. 2007. The CoNLL 2007 Shared Task on Dependency Parsing. EMNLP- CoNLL 2007: 915-932.",
	"links": null
	},
	"BIBREF25": {
	"ref_id": "b25",
	"title": "BioInfer: A corpus for information extraction in the biomedical domain",
	"authors": [
	{
	"first": "S",
	"middle": [],
	"last": "Pyysalo",
	"suffix": ""
	},
	{
	"first": "F",
	"middle": [],
	"last": "Ginter",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Heimonen",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Bj\u00f6rne",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Boberg",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Jarvinen",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Salakoski",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "BMC Bioinformatics",
	"volume": "8",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Pyysalo S., Ginter F., Heimonen J., Bj\u00f6rne J., Boberg J., Jarvinen J. and Salakoski T. 2007. BioInfer: A corpus for information extraction in the biomedical domain. BMC Bioinformatics, 8:50.",
	"links": null
	},
	"BIBREF26": {
	"ref_id": "b26",
	"title": "Exploiting constituent dependencies for tree kernel-based semantic relation extraction",
	"authors": [
	{
	"first": "L",
	"middle": [
	"H"
	],
	"last": "Qian",
	"suffix": ""
	},
	{
	"first": "G",
	"middle": [
	"D"
	],
	"last": "Zhou",
	"suffix": ""
	},
	{
	"first": "Q",
	"middle": [
	"M"
	],
	"last": "Zhu",
	"suffix": ""
	},
	{
	"first": "P",
	"middle": [
	"D"
	],
	"last": "Qian",
	"suffix": ""
	}
	],
	"year": 2008,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "697--704",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Qian L.H., Zhou G.D., Zhu Q.M. and Qian P.D. 2008. Exploiting constituent dependencies for tree ker- nel-based semantic relation extraction. COLING- 2008: 697-704. Manchester, UK.",
	"links": null
	},
	"BIBREF27": {
	"ref_id": "b27",
	"title": "The CoNLL-2008 Shared Task on joint parsing of syntactic and semantic dependencies",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Surdeanu",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Johansson",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Meyers",
	"suffix": ""
	},
	{
	"first": "L",
	"middle": [],
	"last": "M\u00e0rquez",
	"suffix": ""
	},
	{
	"first": "Nivre",
	"middle": [
	"J"
	],
	"last": "",
	"suffix": ""
	}
	],
	"year": 2008,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Surdeanu M., Johansson R., Meyers A., M\u00e0rquez L., and Nivre J. 2008. The CoNLL-2008 Shared Task on joint parsing of syntactic and semantic depend- encies. CoNLL'2008.",
	"links": null
	},
	"BIBREF28": {
	"ref_id": "b28",
	"title": "Syntactic features for protein-protein interaction extraction",
	"authors": [
	{
	"first": "R",
	"middle": [],
	"last": "Saetre",
	"suffix": ""
	},
	{
	"first": "K",
	"middle": [],
	"last": "Sagae",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Tsujii",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "Proceedings of LBM'07",
	"volume": "",
	"issue": "",
	"pages": "6--7",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Saetre R., Sagae K. and Tsujii J. 2007. Syntactic fea- tures for protein-protein interaction extraction. In Proceedings of LBM'07, (319): 6.1-6.14.",
	"links": null
	},
	"BIBREF29": {
	"ref_id": "b29",
	"title": "A Comprehensive Benchmark of Kernel Methods to Extract Protein-Protein Interactions from Literature",
	"authors": [
	{
	"first": "D",
	"middle": [],
	"last": "Tikk",
	"suffix": ""
	},
	{
	"first": "P",
	"middle": [],
	"last": "Thomas",
	"suffix": ""
	},
	{
	"first": "P",
	"middle": [],
	"last": "Palaga",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Hakenberg",
	"suffix": ""
	},
	{
	"first": "U",
	"middle": [],
	"last": "Leser",
	"suffix": ""
	}
	],
	"year": 2010,
	"venue": "PLoS Computational Biology",
	"volume": "6",
	"issue": "7",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Tikk D., Thomas P., Palaga P., Hakenberg J. and Le- ser U. 2010. A Comprehensive Benchmark of Ker- nel Methods to Extract Protein-Protein Interactions from Literature. PLoS Computational Biology, 6(7).",
	"links": null
	},
	"BIBREF30": {
	"ref_id": "b30",
	"title": "Kernel Methods for Relation Extraction",
	"authors": [
	{
	"first": "D",
	"middle": [],
	"last": "Zelenko",
	"suffix": ""
	},
	{
	"first": "C",
	"middle": [],
	"last": "Aone",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Richardella",
	"suffix": ""
	}
	],
	"year": 2003,
	"venue": "Journal of Machine Learning Research",
	"volume": "",
	"issue": "2",
	"pages": "1083--1106",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Zelenko D., Aone C. and Richardella A.. 2003. Ker- nel Methods for Relation Extraction. Journal of Machine Learning Research, (2): 1083-1106.",
	"links": null
	},
	"BIBREF31": {
	"ref_id": "b31",
	"title": "A Composite Kernel to Extract Relations between Entities with both Flat and Structured Features",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Su",
	"suffix": ""
	},
	{
	"first": "G",
	"middle": [
	"D"
	],
	"last": "Zhou",
	"suffix": ""
	}
	],
	"year": 2006,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Zhang M., Zhang J., Su J. and Zhou G.D. 2006. A Composite Kernel to Extract Relations between Entities with both Flat and Structured Features.",
	"links": null
	},
	"BIBREF32": {
	"ref_id": "b32",
	"title": "Semantic Role Labeling using a Grammar-driven Convolution Tree Kernel",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	},
	{
	"first": "W",
	"middle": [
	"X"
	],
	"last": "Che",
	"suffix": ""
	},
	{
	"first": "G",
	"middle": [
	"D"
	],
	"last": "Zhou",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [
	"T"
	],
	"last": "Aw",
	"suffix": ""
	},
	{
	"first": "C",
	"middle": [
	"L"
	],
	"last": "Tan",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Liu",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [],
	"last": "Li",
	"suffix": ""
	}
	],
	"year": 2008,
	"venue": "IEEE Transaction on Audio, Speech and Language Processing",
	"volume": "16",
	"issue": "7",
	"pages": "1315--1329",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Zhang M., Che W.X., Zhou G.D., Aw A.T., Tan C.L., Liu T. and Li S. 2008. Semantic Role Labeling us- ing a Grammar-driven Convolution Tree Kernel. IEEE Transaction on Audio, Speech and Language Processing. 2008, 16(7): 1315-1329.",
	"links": null
	},
	"BIBREF33": {
	"ref_id": "b33",
	"title": "Extracting relations with integrated information using kernel methods",
	"authors": [
	{
	"first": "S",
	"middle": [
	"B"
	],
	"last": "Zhao",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Grishman",
	"suffix": ""
	}
	],
	"year": 2005,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "419--426",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Zhao S.B. and Grishman R. 2005. Extracting relations with integrated information using kernel methods. ACL-2005: 419-426.",
	"links": null
	},
	"BIBREF34": {
	"ref_id": "b34",
	"title": "Exploring various knowledge in relation extraction",
	"authors": [
	{
	"first": "G",
	"middle": [
	"D"
	],
	"last": "Zhou",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Su",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	}
	],
	"year": 2005,
	"venue": "Proceedings of ACL'2005",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Zhou G.D., Su J., Zhang J. and Zhang M. 2005. Ex- ploring various knowledge in relation extraction. In Proceedings of ACL'2005: 427-434. 25-30 June 2005, Ann Arbor, Michigan, USA",
	"links": null
	},
	"BIBREF35": {
	"ref_id": "b35",
	"title": "Tree Kernel-based Relation Extraction with Context-Sensitive Structured Parse Tree Information",
	"authors": [
	{
	"first": "G",
	"middle": [
	"D"
	],
	"last": "Zhou",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [
	"H"
	],
	"last": "Ji",
	"suffix": ""
	},
	{
	"first": "Q",
	"middle": [
	"M"
	],
	"last": "Zhu",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "728--736",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Zhou G.D., Zhang M., Ji D.H. and Zhu Q.M. 2007a. Tree Kernel-based Relation Extraction with Con- text-Sensitive Structured Parse Tree Information. EMNLP/CoNLL-2007: 728-736.",
	"links": null
	},
	"BIBREF36": {
	"ref_id": "b36",
	"title": "Extracting relation information from text documents by exploring various types of knowledge. Information Processing and Management",
	"authors": [
	{
	"first": "G",
	"middle": [
	"D"
	],
	"last": "Zhou",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "",
	"volume": "43",
	"issue": "",
	"pages": "969--982",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Zhou G.D. and Zhang M. 2007b. Extracting relation information from text documents by exploring various types of knowledge. Information Process- ing and Management. 43(2007): 969-982.",
	"links": null
	}
	},
	"ref_entries": {
	"FIGREF1": {
	"uris": null,
	"text": "Procedure for generating SDP-CPT Figure 3. Generation of an example of SDP-CPT",
	"type_str": "figure",
	"num": null
	}
	}
	}
	}